Large-scale building instance segmentation in diverse urban scenes based on LiDAR data
Year: 2026
Type: article
Abstract: Building 3D profiles constitute critical factors in urban environmental analyses, including assessments of urban heat island effects, carbon emissions, population distribution, and related challenges. LiDAR data provide comprehensive high-resolution 3D information, enabling the acquisition of detailed 3D building profiles. However, in dense urban environments, buildings are in diverse shapes and heights, and are close to one another, making it challenging to extract building instances. In this paper, we propose BuildGroup, a novel building instance segmentation model for both Airborne Laser Scanning (ALS)-based and Unmanned Aerial Vehicles (UAV) datasets. It employs the progressive learning strategy in the feature aggregation, which uses prior 2D knowledge from the Segment Anything Model (SAM) to infer spatial relationships within point clouds and mitigate cross-source discrepancies. An offset enhancement module is introduced to mitigate the effects of diverse point densities, improving instance clustering by compacting point cloud distributions. Additionally, an edge-enhanced loss function prioritizes boundary point accuracy during training to enhance morphological precision. Furthermore, we present a building LiDAR dataset with instance-level building information, which fills a key gap in real-world, large-scale urban point cloud segmentation, particularly for ALS-based sparse façade data. Experimental results validate the framework’s capability to produce accurate building instance information across varying urban densities. All code and data are available at https://github.com/littlebeen/BuildGroup .
Institutions: University of Hong Kong, Sun Yat-sen University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,550
OpenAlex ID: W7223131292